CoBiTS: Single-detector discrimination of binary black hole signals from glitches using deep learning

TL;DR

CoBiTS addresses the challenge of distinguishing BBH gravitational-wave signals from non-Gaussian glitches in single-detector data to enable fast, low-latency vetting. The method uses a Conformer-based encoder to process raw strain in sliding windows and outputs two logits representing the presence of a BBH signal and a glitch, with ECDF-based calibration for false-alarm control. It introduces a three-part architecture (MHFA, Conformer encoder, learned pooling) and demonstrates robust discrimination even when a signal overlaps a glitch, outperforming a ResNet baseline and approaching matched-filtering performance in the absence of glitches. The work shows strong generalization to real events and provides explainability through input attribution, offering practical value for rapid follow-up and alert-generation in the upcoming observing runs.

Abstract

We develop a Conformer neural network, called Conformer Binary neTwork Search, or CoBiTS, for distinguishing binary black hole (BBH) gravitational wave (GW) signals from non-Gaussian and non-stationary noise artifacts in the data from current generation LIGO-Virgo-KAGRA detectors. A large subset of these transient noise artifacts, termed as ``glitches'' for short, trigger BBH search templates. Some of them go on to produce detection candidates and require human vetting, supported by data quality tools, to be correctly identified and vetoed. In its current version, CoBiTS takes as inputs single-detector strain timeseries snippets, claimed by other search pipelines to be containing GW candidates, and outputs the significance of each snippet to contain a BBH signal and a glitch. CoBiTS is shown to be particularly effective in discriminating high-mass BBH signals from blips and scattered light glitches, even when a signal is near concurrent or overlapping with a glitch. The performance of CoBiTS gains from employing Conformer, which is a specialized model that combines convolutional layers and Transformer architecture for sequence modeling tasks. Conformer is especially good at leveraging the strengths of both convolutional layers -- for local feature extraction -- and self-attention layers -- for capturing long-range dependencies.

Figures (13)

• Figure 1: Model diagram/flowchart of CoBiTS. The network inputs timeseries data of a certain duration ($T$) from a single (1) channel, e.g., in the form of a GW strain time-series from LIGO-Hanford. Here, $[1, \mathrm{T}]$ defines the dimensions of that dataset. That input is passed through the Multi-Head Feature Extractor (MHFA), which comprises three parallel convolutional heads, each consisting of three convolutional layers. The kernel size of each layer is located within the curly braces in the diagram. The first layer of each head expands the input to $[\mathrm{d}_\mathrm{model} / 3, *]$, and the following two layers consist of many-to-many operations preserving this channel dimension. After these convolutional heads, the resultant features are concatenated along the channel dimension to form outputs of shape $[\mathrm{d}_\mathrm{model},\mathrm{S}]$. Second, after a learned positional embedding is added to the data, it passes through six conformer encoder layers. These encoder layers preserve the shape of the data, but process it further while modeling both global and local temporal dependencies. Finally, these attended features are pooled along the temporal dimension, and, through a simple feed-forward dense network, reduced to the two output logits. See Appendix \ref{['subsec:model_arch']}.
• Figure 2: Plot of the empirical cumulative distribution function (ECDF) generated from the 2000 background samples in the test set. The grey shaded region is the 95% confidence interval from the two-sided Dvoretzky-Kiefer-Wolfowitz-Massart (DKWM) massart1990 inequality.
• Figure 3: Plot of model output logit versus matched filtering SNR over the entire test set. The vertical dotted line is at an SNR of 8.
• Figure 4: ROC curves of different search methods for BBH signals with and without glitches in real GW data. In all cases, CoBiTS outperforms ResNet most likely due to its ability to discriminate based on both local and global attention. All three models perform nearly equally well in the case where the data contains only BBH signals and no glitches.
• Figure 5: Detection efficiencies for various parameter bins -- for CoBiTS and a standard matched-filtering SNR search, which employed the $\chi^2$ statistic as described in Ref. nitz_distinguishing_2018. The threshold for detection in this plot is $\mathrm{FAP} < 10^{-3}.$